linear programming simplex algorithm, duality and dual...

$: Linear programming simplex algorithm, duality and dual ...artax.karlin.mff.cuni.cz/~branm1am/.../01_Branda_DSimplex_2017.pdf · Linear programming { simplex algorithm, duality and$
Linear programming – simplex algorithm, duality anddual simplex algorithm

Martin Branda

Charles University in PragueFaculty of Mathematics and Physics

Department of Probability and Mathematical Statistics

Computational Aspects of Optimization

2017-02-27 1 / 40

Linear programming

Content

1 Linear programming

2 Primal simplex algorithm

3 Duality in linear programming

4 Dual simplex algorithm

5 Software tools for LP

2017-02-27 2 / 40

Linear programming

Linear programming

Standard form LP

min cT x

s.t. Ax = b,

x ≥ 0.

A ∈ Rm×n, h(A) = h(A|b) = m.

M = {x ∈ Rn : Ax = b, x ≥ 0}.

2017-02-27 3 / 40

Linear programming

Linear programming

Decomposition of M:

Convex polyhedron P – uniquely determined by its vertices (convexhull)

Convex polyhedral cone K – generated by extreme directions(positive hull)

Direct method (evaluate all vertices and extreme directions, compute thevalues of the objective function ...)

2017-02-27 4 / 40

Linear programming

Linear programming trichotomy

One of these cases is valid:

1. M = ∅2. M 6= ∅: the problem is unbounded

3. M 6= ∅: the problem has an optimal solution (at least one of thesolutions is vertex)

2017-02-27 5 / 40

Primal simplex algorithm

Content






2017-02-27 6 / 40


1914–20052017-02-27 7 / 40


Simplex algorithm – basis

Basis B = regular square submatrix of A, i.e.

A = (B|N).

We also consider B = {i1, . . . , im}.We split the objective coefficients and the decision vector accordingly:

cT = (cTB , cTN ),

xT (B) = (xTB (B), xTN (B)),

whereB · xB(B) = b, xN(B) ≡ 0.

Feasible basis, optimal basis.

Basic solution(s).

2017-02-27 8 / 40


Simplex algorithm – simplex table

xT

cT

cB xB(B) B−1b B−1A

cTB B−1b cTB B−1A− cT

2017-02-27 9 / 40


Simplex algorithm – simplex table

Feasibility condition:B−1b ≥ 0.

Optimality condition:

cTB B−1A− cT ≤ 0.

2017-02-27 10 / 40


Simplex algorithm – a step

If the optimality condition is not fulfilled:

Denote the criterion row by

δT = cTB B−1A− cT .

Find δi > 0 and denote the corresponding column by

ρ = B−1A•,i ,

where A•,i is the i−th column of A.

Minimize the ratios

u = arg min

{xu(B)

ρu: ρu > 0, u ∈ B

}.

Substitute xu by xi in the basic variables, i.e. B = B \ {u} ∪ {i}.

2017-02-27 11 / 40



Denote by B the new basis. Define a direction

∆u = −ρu, u ∈ B,

∆i = 1,

∆j = 0, j /∈ B ∪ {i}.

If ρ ≤ 0 (u = ∅), then the problem is unbounded (cT x → −∞).Otherwise, we can move from the current basic solution to anotherone

x(B) = x(B) + t∆,

where 0 ≤ t ≤ xu(B)ρu

. We should prove that the new solution is a feasiblebasic solution and that the objective value decreases ...

2017-02-27 12 / 40



New solution is feasible:

x(B) ≥ 0,

Ax(B) = Ax(B) + tA∆

= Ax(B)− tBρ+ tA•,i

= b − tBB−1A•,i + tA•,i = b.

Objective value decreases

cT x(B) = cT x(B) + tcT∆

= cT x(B)− tcTB ρ+ tci

= cT x(B)− t(cTB B−1A•,i − ci )

= cT x(B)− tδi ,

where δi > 0 is the element of the criterion row.

2017-02-27 13 / 40



If ρ ≤ 0, then x(B) is feasible for all t ≥ 0 and the objective valuedecreases in the direction ∆.

Otherwise the step length t is bounded by xu(B)ρu

. In this case, the new

basis B is regular, because we interchange one unit vector by anotherone using the column i with ρu > 0 element (on the right position).

2017-02-27 14 / 40


Simplex algorithm – pivot rules

Rules for selecting the entering variable if there are several possibilities:

Largest coefficient in the objective function

Largest decrease of the objective function

Steepest edge – choose an improving variable whose entering intothe basis moves the current basic feasible solution in a directionclosest to the direction of the vector c

maxcT (xnew − xold)

‖xnew − xold‖.

Computationally the most successful.

Blands’s rule – choose the improving variable with the smallestindex, and if there are several possibilities for the leaving variable, alsotake the one with the smallest index (prevents cycling)

Matousek and Gartner (2007).

2017-02-27 15 / 40


Simplex algorithm – example

3 -1 0 0

x1 x2 x3 x40 x3 2 -2 1 1 00 x4 1 1 -2 0 1

0 -3 1 0 0

-1 x2 2 -2 1 1 04 x4 5 -3 0 2 1

-2 -1 0 -1 0

Moving in direction ∆T = (0, 1,−1, 2), i.e.

(0, 2, 0, 5) = (0, 0, 2, 1) + t · (0, 1,−1, 2),

where t = 2.

2017-02-27 16 / 40


Simplex algorithm – unbounded problem

-2 -1 0 0

x1 x2 x3 x40 x3 2 -2 1 1 00 x4 1 1 -2 0 1

0 2 1 0 0

-1 x2 2 -2 1 1 00 x4 5 -3 0 2 1

-2 4 0 -1 0

Unbounded in direction ∆T = (1, 2, 0, 3).

2017-02-27 17 / 40

Duality in linear programming

Content






2017-02-27 18 / 40


Linear programming duality

Primal problem

(P) min cT x

s.t. Ax ≥ b,

x ≥ 0.

and corresponding dual problem

(D) max bT y

s.t. AT y ≤ c ,

y ≥ 0.

2017-02-27 19 / 40



Denote

M = {x ∈ Rn : Ax ≥ b, x ≥ 0},N = {y ∈ Rm : AT y ≤ c, y ≥ 0},

Weak duality theorem:

bT y ≤ cT x , ∀x ∈ M,∀y ∈ N.

Equality holds if and only if (iff) complementarity slackness conditions arefulfilled:

yT (Ax − b) = 0,

xT (AT y − c) = 0.

2017-02-27 20 / 40



Duality theorem: If M 6= ∅ and N 6= ∅, than the problems (P), (D)have optimal solutions.

Strong duality theorem: The problem (P) has an optimal solution ifand only if the dual problem (D) has an optimal solution. If oneproblem has an optimal solution, than the optimal values are equal.

2017-02-27 21 / 40


Duality – production planning

Optimize the production of the following products V1, V2, V3 made frommaterials M1, M2.

V1 V2 V3 Constraints

M1 1 0 2 54 kgM2 2 3 1 30 kg

Gain ($/kg) 10 15 10

2017-02-27 22 / 40


Duality

Primal problem

(P)

max 10x1 + 15x2 + 10x3s.t. x1 + 2x3 ≤ 54,

2x1 + 3x2 + x3 ≤ 30,x1 ≥ 0,

x2 ≥ 0,x3 ≥ 0.

Dual problem

(D)

min 54y1 + 30y2s.t. y1 + 2y2 ≥ 10,

3y2 ≥ 15,2y1 + y2 ≥ 10,y1 ≥ 0,

y2 ≥ 0.

2017-02-27 23 / 40


Duality

Optimal solution of (D) y =(52 , 5)T

.

Using the complementarity slackness conditions x = (0, 1, 27)T .The optimal values (gains) of (P) and (D) are 285.

Both (P) constraints are fulfilled with equality, thus there in nomaterial left.

Dual variables are called shadow prices and represent the prices ofsources (materials).

Sensitivity: If we increase (P) r.h.s. by one, then the objective valueincreases by the shadow price.

The first constraint of (D) is fulfilled with strict inequality with thedifference 2.5 $, called reduced prices, and the first product is notproduced. The producer should increase the gain from V1 by thisamount to become profitable.

2017-02-27 24 / 40


Transportation problem

xij – decision variable: amount transported from i to j

cij – costs for transported unit

ai – capacity

bj – demand

ASS.∑n

i=1 ai ≥∑m

j=1 bj .(Sometimes ai , bj ∈ N.)

2017-02-27 25 / 40



Primal problem

minn∑

i=1

m∑j=1

cijxij

s.t.

m∑j=1

xij ≤ ai , i = 1, . . . , n,

n∑i=1

xij ≥ bj , j = 1, . . . ,m,

xij ≥ 0.

2017-02-27 26 / 40



Dual problem

maxn∑

i=1

aiui +m∑j=1

bjvj

s.t. ui + vj ≤ cij ,

ui ≤ 0,

vj ≥ 0.

Interpretation: −ui price for buying a unit of goods at i , vj price for sellingat j .

2017-02-27 27 / 40



Competition between the transportation company (which minimizes thetransportation costs) and an “agent” (who maximizes the earnings):

n∑i=1

aiui +m∑j=1

bjvj ≤n∑

i=1

m∑j=1

cijxij

2017-02-27 28 / 40



Apply KKT optimality conditions to primal LP ... we will see relationswith NLP duality.

2017-02-27 29 / 40

Dual simplex algorithm

Content






2017-02-27 30 / 40



Primal problem (standard form)

min cT x

s.t. Ax = b,

x ≥ 0.

and corresponding dual problem

max bT y

s.t. AT y ≤ c ,

y ∈ Rm.

2017-02-27 31 / 40



Dual simplex algorithm works with

dual feasible basis B and

basic dual solution y(B),

where

BT y(B) = cB ,

NT y(B) ≤ cN .

2017-02-27 32 / 40



Primal feasibility B−1b ≥ 0 is violated until reaching the optimalsolution.Primal optimality condition is always fulfilled:

cTB B−1A− cT ≤ 0.

Using A = (B|N), cT = (cTB , cTN ), we have

cTB B−1B − cTB = 0,

cTB B−1N − cTN ≤ 0,

Setting y = (B−1)T cB

BT y = cTB ,

NT y ≤ cTN .

Thus, y is a basic dual solution.2017-02-27 33 / 40


Dual simplex algorithm – a step

... uses the same simplex table.

Find index u ∈ B such that xu(B) < 0 and denote the correspondingrow by

τT = (B−1A)u,•.

Denote the criterion row by

δT = cTB B−1A− cT ≤ 0.

Minimize the ratios

i = arg min

{δiτi

: τi < 0

}.

Substitute xu by xi in the basic variables, i.e. B = B \ {u} ∪ {i}. Wemove to another basic dual solution.

2017-02-27 34 / 40


Example – dual simplex algorithm

The problem is dual nondegenerate if for all dual feasible basis B it holds

(AT y(B)− c)j = 0, j ∈ B,

(AT y(B)− c)j < 0, j /∈ B.

If the problem is dual nondegenerate, then the dual simplex algorithm endsafter finitely many steps.

2017-02-27 35 / 40



min 4x1 + 5x2

x1 + 4x2 ≥ 5,

3x1 + 2x2 ≥ 7,

x1, x2 ≥ 0.

2017-02-27 36 / 40



4 5 0 0

x1 x2 x3 x40 x3 -5 -1 -4 1 00 x4 -7 -3 -2 0 1

0 -4 -5 0 0

0 x3 -8/3 0 -10/3 1 -1/34 x1 7/3 1 2/3 0 -1/3

28/3 0 -7/3 0 -4/3

5 x2 8/10 0 1 -3/10 1/104 x1 18/10 1 0 2/10 -4/10

112/10 0 0 -7/10 -11/10

The last solution is primal and dual feasible, thus optimal.

2017-02-27 37 / 40

Software tools for LP

Content






2017-02-27 38 / 40



Matlab

Mathematica

GAMS

Cplex studio

AIMMS

...

R

MS Excel

...

2017-02-27 39 / 40


Literature

Bazaraa, M.S., Sherali, H.D., and Shetty, C.M. (2006). Nonlinear programming:theory and algorithms, Wiley, Singapore, 3rd edition.

Boyd, S., Vandenberghe, L. (2004). Convex Optimization, Cambridge UniversityPress, Cambridge.

P. Lachout (2011). Matematicke programovanı. Skripta k (zanikle) prednasceOptimalizace I (IN CZECH).

Matousek and Gartner (2007). Understanding and using linear programming,Springer.

2017-02-27 40 / 40

linear programming simplex algorithm, duality and dual...

Documents